AC/DC: Alternating Compressed/DeCompressed Training of Deep Neural Networks

We refine the principle of Iterative Hard Thresholding to propose a simple and effective protocol for unstructured sparse training.’

Official NeurIPS version arXiv version